Spaces:

neuralmagic
/

llama-2-sparse-transfer-chat-deepsparse

Sleeping

mgoin commited on Mar 19

Commit

51aa29a

•

1 Parent(s): d4ee8c2

Update app.py

Files changed (1) hide show

app.py CHANGED Viewed

@@ -19,11 +19,12 @@ MAX_MAX_NEW_TOKENS = 1024
 DEFAULT_MAX_NEW_TOKENS = 200
 # Setup the engine
-pipe = deepsparse.Pipeline.create(
     task="text-generation",
     model_path=MODEL_ID,
     sequence_length=MAX_MAX_NEW_TOKENS,
-    prompt_sequence_length=16,
     num_cores=8,
 )

 DEFAULT_MAX_NEW_TOKENS = 200
 # Setup the engine
+from deepsparse.legacy import Pipeline
+pipe = Pipeline.create(
     task="text-generation",
     model_path=MODEL_ID,
     sequence_length=MAX_MAX_NEW_TOKENS,
+    prompt_sequence_length=8,
     num_cores=8,
 )